Quantum-Driven Chaos-Informed Deep Learning Framework for Efficient Feature Selection and Intrusion Detection in IoT Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors proposed a quantum-driven, chaos-informed deep learning framework for efficient feature selection and intrusion detection in IoT networks, addressing the limitations of traditional IDS (high false positives, computational inefficiency, and poor scalability). The hybrid model integrates quantum computing, chaos theory, and deep learning through four novel modules: Chaotic swarm intelligence (for global feature optimization), Quantum diffusion modeling (probabilistic feature refinement), Transformer-guided ranking (dynamic feature prioritization), and Multi-agent reinforcement learning (adaptive policy optimization).

The framework achieves 75% feature reduction, 4% higher classification accuracy, and 40% lower computational overhead compared to state-of-the-art models, as validated on NSL-KDD, CICIDS2017, and UNSW-NB15 datasets. Since the public datasets are used in the evaluation, it will be easier for readers to replicate and compare the results. 

While the proposed framework can lead to benefits such as feature reduction while improving accuracy, it is unclear on performance improvement brought by each component of the framework. Is the combination of components optimal? The author should provide further 
experiments for the analysis.

The writing can be improved. For example, the math notation on mutual information is unclear to me. Figure 3 should be presented in a more clear way.  

Author Response

Responses to the comments/suggestions by the Reviewer-1

Comment-1:

While the proposed framework can lead to benefits such as feature reduction while improving accuracy, it is unclear the performance improvement brought by each component of the framework. Is the combination of components optimal? The author should provide further
experiments for the analysis.

Response:

Thank you for pointing this out. We agree with this comment. Therefore, we have added further experiments in the form of an ablation study. which shows the performance improvements of each component in the framework. [Section 4: page-15 last-para, Page-16 Table-10 3^rd-para]

Comment-2:

The writing can be improved. For example, the math notation on mutual information is unclear to me. Figure 3 should be presented more clearly.

Response:

We agree to the reviewer’s suggestion. We have, revised the mutual information accordingly; specifically, the fitness function is expressed as Equation 3. This formulation ensures that features with strong discriminative capacity are prioritized while penalizing redundancy. Likewise, Figure 3 has also been reformatted to highlight performance metrics across datasets with clearer axes, color-coded markers, and enlarged annotations, improving the interpretability of the comparative analysis. [Section 3: Page-5 Equation-3 4^th-para, Page-14 Figure-3]

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for your submission. I find the paper an interesting read, while having the following concerns.

- The work targets IDS for IoT scenario. I'm wondering what's the exact threat model for these attacks? Since IoT networks are mentioned throughout, I'm wondering if it's a specific network or the solutions shall be different accordingly. The target system will also decide where the solution is implemented.

- Regarding the related works, the final two paragraphs are discussing some advanced techniques, but no references are provided. Given this, the evaluation also only compares against some baselines that are not using the advanced techniques. It is not clear how the proposed solution is superior compared to state-of-the-art.

- The proposed technique seems interesting, but why it works better is largely unclear. The authors' claim is the existing solutions would "entrapped in local optima", but I think those deep  learning based approach won't face the same issue. On the other hand, my feeling is that the proposed approach seems to have the benefit of achieving good performance while using less overhead. However, this needs to be justified as well - from my understanding, the training is just one-time thing and the inference won't be of high overhead.

Author Response

Responses to the comments/suggestions by the Reviewer-2

Comment-1:

The work targets IDS for IoT scenario. I'm wondering what's the exact threat model for these attacks? Since IoT networks are mentioned throughout, I'm wondering if it's a specific network or the solutions shall be different accordingly. The target system will also decide where the solution is implemented.

 

Response:

We thank the reviewer for pointing this out. We revised the manuscript by specifying the target system where this solution can be implemented. [Section 1: Page-1 1^st para up to line-8]

Comment-2:

Regarding the related works, the final two paragraphs discuss some advanced techniques, but no references are provided. Given this, the evaluation also only compares against some baselines that do not use the advanced techniques. It is not clear how the proposed solution is superior compared to the state-of-the-art.

Response:

We agreed to point out that. We have revised the related works section by referring to advanced techniques, and we have compared the evaluation with them. [Section 2: Page-4 3^rd para]

Comment-3:

The proposed technique seems interesting, but why it works better is largely unclear. The authors' claim is that the existing solutions would be "entrapped in local optima", but I think those deep learning based approaches won’t face the same issue. On the other hand, my feeling is that the proposed approach seems to have the benefit of achieving good performance while using less overhead. However, this needs to be justified as well - from my understanding, the training is just a one-time thing and the inference won't be of high overhead.

Response:

We thank the reviewer for pointing this out. Classic deep learning IDSs with raw or high-dimensional features require long training and memory usage. Chaotic swarm intelligence and quantum diffusion preserve only the most useful data before categorization, halving training complexity. Retraining deep models is necessary for real-world IoT IDS to respond to changing threat signatures and network behaviors. This scenario requires reducing feature space and computational expense. For edge-level deployments, feature optimization and quantum-modulated graph attention increase generalization under limited retraining and low inference latency in the process. We have revised the paper by specifying the same in Section 4. [Section 4: Page-17 2^nd para]

Rewritten the “Conclusions & Future Scopes” section as per the suggestion.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for submitting your revision. I'm generally happy with the new version but have three suggestions:

1. Please include some citations in the first paragraph of intro to highlight the possible attacks against IoT;
2. It's a bit strange to have the last paragraph in section 2 (line 166-179). Section 2 is background, while this paragraph seems to be the overview of the proposed work. Do we need it? Instead, probably summarize the deficiencies of existing works, which would lead to the authors' design.
3. In section 3, before introducing the actual techniques, some high level description of why the proposed approach works better would be beneficial. Specifically, if the authors follow my suggestion 2, then here you could mention why the existing deficiencies would be addressed by the solution.

Author Response

Reviewer-2-Round-2

Thanks to Reviewer-2 for suggesting minor revision to our revised manuscript.

Comment-1: Please include some citations in the first paragraph of intro to highlight the possible attacks against IoT.

Response: We thank the reviewer for the valuable suggestion. We revised the manuscript by giving the citation to the specified attacks in the first paragraph of Introduction Section. [Section 1: Page-1 1^st para line no 23]

Comment-2: It's a bit strange to have the last paragraph in section 2 (line 166-179). Section 2 is background, while this paragraph seems to be the overview of the proposed work. Do we need it? Instead, probably summarize the deficiencies of existing works, which would lead to the authors' design.

Response: We agreed to point out that. We have revised the related works section by restructuring last paragraph which is just summarizing the limitations of existing works. [Section 2: Page-4 last para (line 166-173)]

Comment-3: In section 3, before introducing the actual techniques, some high-level description of why the proposed approach works better would be beneficial. Specifically, if the authors follow my suggestion 2, then here you could mention why the existing deficiencies would be addressed by the solution.

Response: We thank the reviewer for pointing this out. We revised the proposed methodology section by adding some high-level description of why the proposed approach works better.  [Section 4: 1^st para (line 175- 184]

We hope with this revision our manuscript will be eligible for the poaaible inclusion in the journal.

Thank you,

Authors